Research
TraceForge generates reasoning traces for distillation without access to hidden teacher reasoning
The method trains a generator and critic to write reasoning traces in a student model's family style, avoiding terms-of-service issues with encrypted reasoning. On coding, it lifts HumanEval+ by 3.7 points and MBPP+ by 7.1 at the final checkpoint in tests with a fixed 5,000-question set. Authentici…
Every outlet that ran it 1 article
Hugging Face PapersResearchLead